3 ms·
The biggest openai news today is not a store that launches next week. It is them turning off API access to their old top of the line models like text-davinci-00
by superkuh 3y ago
The biggest openai news today is not a store that launches next week. It is them turning off API access to their old top of the line models like text-davinci-003 (gpt3.5) and many others (ada, babbage,curie) on Jan 4th (today). https://platform.openai.com/docs/deprecations/deprecation-history https://platform.openai.com/docs/deprecations/deprecation-hi...
Their suggested replacement is a gpt-3.5-turbo-instruct model and those pale in comparison to text-davinci-003 original. It's a shame. Some of my services will stop working (in the sense that users won't care to use the service). The local LLM replacements like Mixtral-8x7B are not as good as text-davinci-003 but they're definitely better than a gpt-3.5-turbo.
- Traubenfuchs 3y agoI wonder what keeps them from offering superior "legacy" models? They could just say those models are less cost/efficiency optimized and raise the costs to cover the expenses + even an extra margin.
- superkuh 3y agoLegal butt-covering in two forms, I'd guess. They said a lot in detail about early models training data sources that could be used against them for spurrious but costly legal attacks. And the old models aren't quite as heavily aligned/censored as the new ones.
- mensetmanusman 3y agoThis is what an undersupply of GPUs feels like.
- infecto 3y agoWhile the latest interation of 3.5-turbo has had some regressions in quality, I have found it to be still very powerful compared to text-davinici-003. Are you sure you cannot tweak more to get it working? I am curious to what kind of task load is benefitting greatly from text-davinci-003?